Orbital Data Centers Hit Physical Barriers in Cooling and Cost

At NVIDIA’s GTC conference, CEO Jensen Huang declared, “Space computing, the final frontier, has arrived.” SpaceX, Google, and startup Starcloud have successively announced plans to build orbital data center constellations comprising thousands of satellites. The vision involves equipping each satellite with AI GPUs, interconnecting them via optical links, and communicating with the ground through microwave links.

Supporters cite advantages of space computing: abundant solar energy, free cooling systems, and immunity to ground-based disruptions such as earthquakes, floods, and protests. However, a detailed examination of physical principles reveals that realizing orbital data centers is far more challenging than Silicon Valley imagines. Based on an analysis from IEEE Spectrum, we delve into these challenges.

The Biggest Misconception: Cooling in Space

“Free cooling” in space is one of the most deeply-rooted misconceptions. While space is indeed frigid, the absence of an atmosphere renders conduction and convection—the primary heat dissipation mechanisms—ineffective. Air cooling and water cooling systems used in terrestrial data centers are completely useless in space.

The only available heat dissipation mechanism in space is releasing heat via radiation. However, preventing chips from overheating requires vast and expensive surface area. AI GPUs generate hundreds of watts of heat. To radiate this heat into space, satellites must be equipped with enormous radiators (radiative coolers).

The area of the radiator depends on the chip’s power consumption and allowable operating temperature. While a terrestrial data center can cool tens of kilowatts per rack, achieving the same cooling capacity in space may require radiators several to dozens of times the satellite’s own surface area. These radiators increase launch costs and significantly add to the satellite’s weight and volume.

The Solar Power and Attitude Control Dilemma

Abundant solar energy also presents complex challenges. Keeping a satellite constantly facing the sun requires sophisticated attitude control systems. Solar panels must track the sun for maximum efficiency, while simultaneously, the radiator must face deep space (the cryogenic background) to release heat.

This creates a conflicting attitude requirement: part of the satellite must always face the sun, and another part must always face space. Meeting this requirement necessitates either a combination of movable solar panels and fixed radiators, or a complex mechanism that rotates the entire satellite. Both options negatively impact cost, weight, and reliability.

Space Radiation and Performance Degradation

Space contains far higher levels of radiation (cosmic rays, solar flares, high-energy particles in the Van Allen belts) than on Earth. This radiation degrades solar panel efficiency, deteriorates radiator surface properties, and—most critically—severely affects the semiconductor chips of AI GPUs.

Semiconductor devices are susceptible to single-event upsets (SEUs) and cumulative total dose effects, leading to malfunctions, performance degradation, and eventual failure. GPUs used in orbital data centers would require radiation-hardened (rad-hard) design, which tends to be orders of magnitude more expensive and lower-performing than commercial off-the-shelf components.

Current AI GPUs are optimized for terrestrial data centers and are not designed for the space environment. Developing dedicated chips with radiation tolerance would require years and massive investment.

Maintenance and Redundancy Challenges

Maintenance in space is extremely difficult. If a satellite fails, the cost of dispatching a spacecraft for repair far exceeds the satellite’s manufacturing cost. Therefore, satellites must carry redundant systems (backup GPUs, power supplies, communication equipment, etc.). This further increases satellite weight and cost.

Redundancy requirements not only drive up launch costs but also increase the complexity of software updates and reconfiguration in orbit. AI workloads are frequently updated, requiring a mechanism to remotely update satellite software. Considering communication latency and bandwidth constraints, operational complexity is orders of magnitude greater than on the ground.

The Reality Shown by Cost Comparison

A rough cost comparison by IEEE Spectrum indicates that operating an AI GPU in space for one year is at least an order of magnitude more expensive than in a terrestrial data center. This estimate includes launch costs, satellite manufacturing, rad-hard component premiums, redundancy systems, and operational/monitoring costs.

While specific figures are not disclosed, launch costs are thousands of dollars per kilogram (approximately $2,700/kg for Falcon 9, with Starship expected to lower this further), whereas terrestrial data center capital expenditure is on the order of a few dollars per watt. Launching a single AI GPU (a few kilograms) into space could cost as much as operating comparable computing power on the ground for several years.

Potential in Specific Niches

Orbital data centers are not entirely without merit. They hold value in areas requiring data processing in remote terrestrial locations (oceans, deserts, polar regions), low-latency Earth observation data processing, and military/security applications where ground infrastructure is weak.

Furthermore, the combination of solar power and radiative cooling theoretically enables 24/7 operation. However, considering economic feasibility, it is reasonable to conclude that, at least with current technology, orbital data centers will be limited to niche applications.

Editorial Opinion

In the short term, widespread recognition of this analysis may dampen excessive investment fervor in space data centers. SpaceX, Google, and startups will be pressed to demonstrate realistic milestones. Specifically, concrete technical demonstrations addressing cooling and radiation tolerance will be required. If small-scale orbital experiments are conducted around 2026-2027, their results will become critical turning points for the industry’s direction.

From a long-term perspective, for space data centers to become economically viable, further reductions in launch costs (practical operation of fully reusable rockets like Starship) and the development of high-performance, radiation-tolerant semiconductors are essential. As semiconductor processes shrink below 3nm, radiation effects become relatively larger. The key will be whether innovative designs can overcome this trade-off—e.g., enhanced error correction codes or hardware-level tolerance improvements.

An editorial question: Is the space data center concept merely a product of “Silicon Valley’s sci-fi optimism,” or is it part of a technology roadmap worthy of long-term investment? Alternative scenarios that cleverly utilize, rather than defy, physical laws—such as construction on the Moon or asteroids, or distributed computing leveraging resources within the solar system—are also worth considering. As terrestrial data centers’ energy consumption and environmental impact grow, space computing may not be the “final frontier” but merely another form of the “battle against cost.”

References

• Why Orbital Data Centers Are Harder Than Silicon Valley Thinks — IEEE Spectrum, published 2026
• Physical Limits of Space Computing — Solidot, published 2026-06-13